文 | 硅谷101
回顾9月，AI圈的重磅新闻一个接一个。
9月4日，OpenAI发布GPT-6 Astra。这个发布报告里只字未提机器人的通用模型，转身就空降了机器人评测榜首，甚至让一批机器人Benchmark直接失效。有人半开玩笑地说：机器人领域等了这么多年的“GPT时刻”，居然是GPT自己。
Meta Muse给每个用户配了一台云端电脑，两周就冲上美区App Store榜首；竞争对手Instinct一个月内估值从25亿涨到100亿美元。个人Agent掀起了新一轮烧钱获客大战，可它们能打开的网页，正在一个一个被关上。
AI for Science方面，Claude只用了21小时就找到一种结构像CRISPR的新酶系统ART，AI圈很兴奋，因为AI开始做科学发现了，生物圈的反应则相对平淡，大家想在实验室里验证了再说，为什么会有这种反差现象？
当模型进化到普通人难以感受到它的进化，研究者的感受是什么样的？本期《硅谷101》，我们请来三位分别站在Agent、机器人和AI制药一线的研究者——Pokee.ai创始人兼CEO朱哲清，破壳机器人创始人、清华大学交叉信息研究院助理教授许华哲，华盛顿大学计算机学院助理教授王晟，一起聊清楚三件事：个人Agent商业化卡在哪？机器人怎么被Astra这样的底层模型冲击？AI制药真正的瓶颈是什么？
以下是这次对话内容的精选：
泓君：Astra发布后，你们有没有使用过这个产品？或者说我们来看网上的demo，大家觉得最让你惊艳的一个demo是什么？
许华哲：因为我是做机器人的，我们测了很多Astra在做机器人任务上面的一些能力。我们看到的第一个点就是，它的适应能力非常强。比如说给它某一个机器人本体，让它去做相应的事情，那它很快会发现，这个机器人的运动速度比较快，或者这个机器人的运动空间是怎样的，比如它撞到桌子了，它会说，这个桌子蛮高的，那我是不是应该把手抬得更高一点？这其实是把CoT（思维链）这种反思过程的能力，变成一个它实际的动作的能力，这是过去的很多机器人的模型和大模型都不具备的。
另一个是它自我更新、自我升级的能力。因为它有这样的反思，它可以一次一次做同样的事情，每一次做得更好。这个其实在网上大家也看到了，有人拿Astra去操作一个机器人机械臂去画金门大桥，那它第一次画的还是比较差的，但是它画完了以后，它会再想怎么样能画的更好，经过几轮的迭代，它就可以画的非常逼真。
我们试了非常多其他的，比如插花，这是一个经典的机器人任务。我们给它加了一些额外难度，那朵花的茎被劈开两半，有点分叉。这个时候它会知道，因为它分叉了，所以我要捏住这个分叉，把它捏窄一点，再往里塞。这样的能力其实还是比较惊艳的。
泓君：它其实已经理解了一些物理世界操作的规范。
许华哲：我觉得本质上还是理解3D，因为这件事情它还是从三维尺寸层面就可以做出一些理解，但是确实这样的理解已经能够帮助到非常多的非动态的物理世界操作任务了。
泓君：我看网上还有一些很有意思的demo是，他们在用Astra做生成3D的视频，非常的快，然后渲染也很好。
许华哲：这是另一个Astra在机器人方面很令人震惊的能力，就是给它几张图片，它可以很快把图片仿真生成出来。学术圈可能把它叫做Real-to-Sim。这件事过去是需要拍大量的视频照片，需要很多人的手工标定修改。但现在你给Astra这样一个指令，它很快就把整个实验室或者办公室的环境给你复现出来了。那你拿着这样的仿真，可以去训你的机器人，那也会有一个在仿真上面的提升。
朱哲清：我们是做模型公司的，Astra在发布的时候，非常强调他们在后训练、模型开发等等上面的一些迭代能力。试下来以后，并没有那么惊艳，但是它确实在长程任务上面和迭代上面，比之前的几版模型有一个比较稳定的提升。与原来的模型相比，它能够对上下文以及之前完成的任务有更稳定的认知，不会出现太多的幻觉的情况。所以从我们的目前的体验来说还是比较好的。
但可能有几个问题：速度相对来说会慢一些，也有可能跟现在用量很高有关系。同时在更加长程的任务上，比如说让它跑一个晚上，幻觉还是会发生。所以它并没有根本上解决，但总体来说还是提升了不少。
我有朋友同测了一些前端以及设计的任务。从前端的审美跟认知上面，我觉得还是有比较大的提升的。可能是在训练的时候，对于设计和审美方向有特殊的训练有关系。
泓君：Bill，这轮Astra的发布，你觉得从模型对行业的影响来看，它跟当年opus 4.6、4.7的发布，从Coding能力的进化来比，它是一个大的节点吗？
朱哲清：单纯Coding来说，可能没有那么大的一个节点。4.5到4.6，当时我们的体感是一个几乎范式式的转变，原来只能依赖它做短程Coding，几个file的改动，做一些简单的小项目，到4.6那一次模型以后，基本上你会感受到工具调用几乎不会出错了，它去做一些长程的Coding任务，可以自主进行了。
而现在Astra只是说可以做到更长程的任务，幻觉是可以降更低，审美变得更好。但它没有做到说，我能够解决一个完全以前完全解决不了的问题。
王晟：我觉得是一个大的节点。我们现在已经都从之前Claude等系列模型切换到GPT-6 Astra来做我们的biomedical内部的任务了。我主要是做AI for Science，里面核心的就是AI for一些biomedical的应用。
我们用Astra有两个特别好用的地方，一个是Computer Use，一个是一些长程的任务。
Computer Use这方面，我们会用Astra来操作一些生物医学领域很多年前编写的那些软件，这些软件在领域内默认要使用，但它们其实都非常难用，界面设计的也不好，还有很多的术语，所以对一些跨领域的人来说，使用成本是非常高的。
但是Astra它的Computer Use功能，可以直接根据一些简单的prompt，就可以操作得比较好。并且我还可以跟它学习怎么用这个软件，毕竟在生物医学领域，很多人对AI工具的接受度比较低。Astra可能把这些软件对应的在网上的一些教程，别人之前用过的一些用户日志，一些反馈全部都以文字的形式学进去了，所以它真的理解怎么用这些软件。
另一方面就是长程的任务，我们生物医学这块，任务会比较复杂，这些任务有时候会持续8至24小时，Astra会做得比较好。它比起其他的一些模型，做得比较慢，但挺让我放心的，说明它可能会更深思熟虑一点。
我们现在一般切换着用，白天用一些反馈比较快的模型，到了准备让模型跑一晚上的时候，就用Astra来设计一个比较长程的pipeline或者prompt。这也是它比较值得信赖的一点，在AI for Science这个领域里面的一些表现。
泓君：我没有想到Computer Use的这个功能居然在AI for Science领域能帮助到大家这么多。
王晟：Astra在这次发布时候做了两个生物医学的demo。其中一个就是Computer Use demo，这个demo在我们圈子内是出圈了，大家发现，原来可以直接用Astra来操作一个生物医学的tool。它当时操作的一个tool叫MultiQC，是一个挺常用的网页工具。我们开始用这个功能，发现它比想象的要好用很多。其实Computer Use这个功能我觉得之前的模型也具备，只是大家没有开始真的思考怎么把它融入我们科研中的pipeline里面。
泓君：Bill，你可不可以一句话解释一下，什么是Computer Use？
朱哲清：Computer Use是通过模型对于视觉的理解，来决定这个agent下一步要在屏幕上点击哪个部分，或者输入哪些文字，来完成一系列的任务。当它可以串联这么几十个动作的时候，可以完成一个最终人类想要达到的一个目的。
泓君：就它跟API的区别是，你不用给接口，不用给密码，它直接像人一样去操作跟使用电脑。
朱哲清：就是它看到的是人类看到的屏幕，看到人类看到的UI，看到人类看到的所有的视觉和可以交互的界面。通过人类可以使用的电脑的输入输出设备来完成这些任务。
泓君：我觉得Computer Use之前好像也有，为什么Astra的发布，又重新把Computer Use带火了？
朱哲清：感觉Computer Use现在比原来又火了一波。之前大家觉得所有的工具都会开API，所有的工具都会有agent access（访问权限）的接口，很多的APP也好，系统也好，都在做agent native（原生智能体）的set up。
但是最近，不管是Instinct也好，还是Meta Muse也好，背后其实都接了浏览器。有一个比较大的问题就是有很多第三方的APP或者Amazon也好，比如说一些shopping的APP也好，travel的一些APP也好，他们其实是不愿意开放API的。
我觉得这又回到了当年Manus的情况，就是Manus最开始做了Browser Use（浏览器操作），后面很多别家也跟上，后面大家发现这个不是很高效。大家都回到Coding和API调用，现在好像有些东西避不开，又回到了Computer Use，所以我觉得最后有可能还是个mixture的状态。
我这边有一个比较有意思的看法，因为我们目前是focus on相对比较小的一些模型，比如说我们基本上现在发布的model都30B左右的模型，是可以在端侧和私有算力上面去进行的。比如说我有一台小的AI box，或者说我有一个AI PC，它就可以跑这个模型了。
未来如果说这个Browser Use本身发生在你的机器上面，而不是发生在云端，那就会改变Browser Use本身可能会给你带来的一个瓶颈。因为现在的远程的Browser Use都是通过商用机器，最多找一个像Bright Data或者说IP代理，假装是消费者用户，去绕开IP校验。像Amazon、Expedia这些网站一般都说，如果是商用的IP，我不会让你进去。
泓君：商家会屏蔽一些AI机器人，不希望它们接入。
朱哲清：奢侈品网站很早以前就会都屏蔽掉。如果用的你们家的网，就是你自己的IP，那就不会再被block了。所以这个有可能会是未来一个破局的趋势，而这件事情有可能会推动很多的app厂商去改变他们的Browser Use机制，也就是说这个browser sandbox（浏览器沙箱）可能是在本地发生，而不是在云端发生。
泓君：你这个观察非常有意思，最近非常火的一个toC的消费级应用，就是Meta Muse，他们的定位是做一个个人的AI助手，它相当于给每个人做了一台云端的虚拟机，你的数据是存在一个在云端属于你的存储里的。根据你刚刚的理论，它其实还是商用的一个IP。它如果未来商业化的话，很有可能是会被商家卡住它这个IP的。
朱哲清：Meta Muse从背后的虚拟机设置来说，其实跟当时的Manus是很像的，它也主打一个Browser Use，因为它很多的使用场景跟旅行和购物有关，在这些use case之下，你是没有办法通过API的调用，来完成这个任务的，基本上都得最后靠浏览器。当时Manus其实走了一模一样的路，被全部block掉了，所以Manus到后期的时候，现在跟Genspark很像，都主打白领工作那一套。
你去看Claude Code、Claude Cowork、ChatGPT Work，最后都往白领工作效率这方面去偏移，核心原因就是因为很多东西一开始能做，但后面很快就被巨头们给全部block掉了。
泓君：比如说亚马逊现在就把Meta Muse给屏蔽掉了，如果有更多的网站去加入进来，它可能能访问的网页跟能操作的东西就越少。
朱哲清：但是他不可能把你电脑上的IP给屏蔽掉，那你自己也上不了亚马逊了，所以这个可能会是一个破局的机会。但是我怀疑这个事情会来回拉扯，在C端的这个agent，目前可能还是不是特别ready。因为做Browser Use，对模型的需求其实挺高的。它很难在短期内拉出一个现象级的体量。
泓君：那你会看好像Meta Muse长远的一个发展吗？
朱哲清：我觉得这波消费者最后一定会land到软硬件结合的状态。我觉得Meta如果能够把wearable（可穿戴设备）做很大，有可能有机会。就是为什么我觉得最后要land到软硬件结合的原因，是因为云端这件事情不能代表这个用户本身，这也是为什么browser把他们给完全block掉的原因，因为一旦换了个云端的虚拟机，它不能够验证说这个人本身就是用户本身。
我觉得用户的ID或者使用agent的那个用户的identity（身份）本身会是这一波toC的一个关键。如果说商家能够去确认这个agent就代表这一个人去做的这么一个操作。有可能很多商家会愿意做这件事情。
但是现在如果用云端的虚拟机，云端的browser，他很有可能是没有办法去验证这件事情，他会觉得说我的流量入口被人剥夺了。比如说我有一个用户，他本来就要访问我，他现在通过agent还是访问我，那其实对于我来说可能是OK的。但如果说原本一个用户，现在搞了20个agent，这20个agent的消费能力跟原来是一样的，这20个agent还不停地比价，不停地在我这边刷流量，把我这边的所有的流量全都弄乱了，对我来说是得不偿失的。
所以我觉得软硬件结合可能会是一个机会，如果说这个硬件本身属于这个用户，他就可以直接确认说这个agent就是这个用户，是一个代表行为，我就会愿意把这个access开给你。
泓君：大家在toB的工作流中会用xAI的Grok Bot吗？
朱哲清：我自己身边用Grok Bot的人更多是在尝鲜，因为它相对比较主动一点。但是你用Codex或者用Claude Code，其实都可以解决这些问题。就真的把它接进工作流的，可能相对少一些。我自己身边好多真正在B端，比如法律、医疗，真正在那些传统的企业里面在工作的人，现在还纠结于怎么能够逃出Microsoft Copilot的泥潭。甚至都不能freely使用Codex和Claude Code，就更别说要到Grok Bot这个级别了。所以我觉得整个行业现在分差非常大，走在前面的人什么都可以随便用，走在后面的人现在还在泥潭里挣扎。
泓君：我最近跟比较前沿的模型的研究员聊啊，听到一个很有意思的观点，他们认为现在在toC跟消费级的agent上，token的使用量其实是花不了多少钱的。现在真正最烧钱的地方，是在OpenAI、Anthropic、DeepMind这几家公司在竞争的核心领域——AI for Science。就是那些未被解决、未被定义的科学问题，他们要去解那些数学题，或者需要跑像王教授这样的一些生物实验，这个是非常非常烧token，而且是内部消耗量最大的一个方向。
那比如说，OpenAI解纳维-斯托克斯方程这一个问题，它内部调用了一个更强的模型，大概用了1万个子agent，发了大概有270万条信息，消耗了约1300亿输出的token。第三方按照零售价给它算了一下，这部分消耗大概是在2200万美元。
因为现在整个模型的进化，我觉得这个对普通人的感知来说已经不大了。但是对前沿的科学家来说，他们对模型能力的提升，感知是非常大的。
朱哲清：这个地方有点想问许教授跟王教授。我自己主要做post training（后训练），focus on小模型多一些，因为我们想做本地私有化的模型。我们能看到的就是，这个层次的模型，在我们优化之后，和frontier model的差距越来越小。
从AI for Science角度，或者说从robotics角度，frontier model在每一代的迭代当中，你们觉得会有多大的提升？
比如说我们从Opus 4.5到4.6的区别，对比4.8到Fable之间的区别，或者说5.6到Astra之间的区别，这个差距可比吗？
王晟：我的观点是，在生物医学领域，它的AI for Science能力确实是有一个很大的提升。但是这个提升主要来自于它的coding能力，这种一般的coding查文献，做长程任务的能力，并不是它对这种生物医学知识能力的提升。有一些benchmark是专门测Biomedical knowledge知识层面的，Astra并没有提升。
我们现在实际用的时候，其实很难去使用Astra，或者Opus、Fable，因为它们对生物医学的任务都会防范的越来越严，它都会refuse to answer，这个是非常严重的问题。所以我们其实内部用的时候都是用Kimi K3或者Muse Spark 1.3。当如果要用Astra的时候，都要用一些比较特殊的方式来跟Astra说，我这不是在做一个危险的任务，我只是在设计一个benchmark等等。所以实际这种大模型的使用体验，随着能力的增强，其实是越来越差的。
还有一个点就是，刚才Bill说的很有道理，其实我也认可，就在各种已经定义的很明确的AI for Science benchmark上，小模型确实是可以和大模型的差距越来越小。但是我个人觉得这些benchmark本身就不是定义的那么好，所以大家追赶的方向，包括大模型进步的方向，都可能不是完全正确的。
这些benchmark定义的时候，大家都会定义一个更好让AI模型去做的，比如说有一个明确的生物数据集，要预测一个明确的问题，这和我们真的做的制药发现其实是gap非常大的。本质上这些现在的生物医学benchmark就是一个Kaggle这种数据科学的比赛，所以呢它就退化成一个MLE-bench（机器学习工程基准）中machine learning的问题。那Astra这种在coding能力上提升，就会在benchmark上有很大的进步。
许华哲：我觉得从Robotics角度，Astra是一个飞跃式的提升。过去无论是Opus 4.5、4.6，还是到Fable的变化，都没有Astra这次给Robotics带来的变化大。因为过去的所有升级，在Robotics的这个测试上面，几乎都没有什么太好的表现，定性上面没有变化。
但Astra现在带来的变化是它整个定性是有变化的，很多的Robotics的benchmark失效了。像RoboDojo，大家相对认可的Robotics尤其是仿真的benchmark，Astra都成了第一名。无论是它用了哪种数据，但是它带来的这个能力，都是一个飞跃式的升级。
但是关于刚刚说的小模型和大模型这个问题，我倒是觉得，长期来说我们一定是需要一个小的，跟Astra能力相当的模型。原因是机器人的任务有时间要求，需要及时响应，需要有高动态任务。比如说我要抖一个毛巾，这样的能力Astra是做不了的，因为它太慢了。它知道我要快动一下，但是它只能写code来做，如果让它直出动作是没有办法去做的。在这种情况下，我们一定需要一个快决策的模型。只不过我们需要把现有的Astra的CoT，或者是它的reasoning的能力，以及3D理解，要把它加进去。
泓君：所以你觉得未来机器人它的方向会变成，大模型去做规划，小模型负责实施控制，这样的分层架构吗？
许华哲：分层架构大家其实一直都还是比较认同的，只不过分层的这一刀，到底切在哪？怎么样的分层方式是大家一直有辩论的。就比如说原来大家普遍认为VLM（视觉语言模型）是做任务规划层，那这个任务规划层大概就是到，我跟你说我要吃西红柿炒蛋，那VLM规划的可能就是打开冰箱门，拿出什么西红柿，拿出鸡蛋等等，底下的小模型或者说具身的模型，它会去执行打开冰箱门究竟怎么办？打鸡蛋怎么做？
但是，Astra这次它最大的一个变化是，它现在能做低层次的动作控制，相当于它给你的指令变成了把手往右伸两厘米，或者把手放在冰箱门的把手上，如果没有抓住的话，他说那你再往前靠一靠……所以它相当于把这个层往下压了。
原来是大家边界很清晰的，但是它现在说我能力非常强，除了慢以外，我可以往下压一点点。那底下的这个就更像是一个反射式的提线木偶。肯定有一些声音说，那是不是未来上层的这个模型可以无限往下压？压到最后是不是就变成一个完全端到端的VLM？Astra直接控制所有事情。
我个人的观点是，我觉得这个还是不是很合理。第一明显是速度，你在网上看到的Astra做控制都是要16倍速或24倍速才能看。第二个，我觉得现在的Astra还是对物理是非常缺乏理解的，尤其是对于人类来说叫直觉，或者叫直觉物理学的那部分。
我随手可能碰一个物体，它如果有水可能会洒，如果是玻璃可能会碎，这些事情可能Astra如果经过很多语言推理是可以理解，但是如果我们希望它去做一个任务当中，立刻有这方面的理解，它可能效果还不够好。这个地方可能还是需要像世界模型，或者是像快速的这个模型，要去做一些直觉式的反应。
泓君：你刚刚提到了它速度慢，它速度慢是因为什么？是因为它的模型太大了吗？
许华哲：模型大肯定是有这样的一个原因，另一个就是它CoT很长，虽然它是加密的CoT，给出来的那个CoT应该是假的，我理解，但是它的内部的CoT是非常长的，所以它做每一步的决策，做的思考是非常冗余的，尤其是面对简单任务的时候，我们观察到面对复杂任务和简单任务，它的思考时长基本上是相似的，这也意味着复杂任务它很聪明，对吧？但是到了简单任务上，它就变成了一个额外开销，使得它看起来非常笨。
泓君：许教授，我知道你们公司的有一个demo，是用筷子夹小球。用你们自己的模型做，跟用Astra做，区别是什么？
许华哲：用我们自己的模型，首先时间上做得快很多；其次，在抓的时候是连续不断地去做尝试的，这个过程中没有太多的逻辑上的思考，我抓了一次失败了，我可以换一个角度去做尝试。
但是如果是Astra的话，它每做一步以后，它都要分析发生了的原因是什么，根据这个原因，它要去给自己指令，相应的我要做什么样的调整。这个过程看起来更像是这个人类的思考过程，但是会导致它很缓慢；第二个，它有时候会陷入到一些僵局里，它有时候无法理解为什么会发生这样的事情，或者是它一旦分析出错了，它就没有办法再进一步去推进了。
还有一个我们看到的点，就是如果我们去抓一个桌面上的物体，旋转这件事情，Astra其实做的相对来说还是弱一些，但是它能做，它要反复尝试怎么样去转，这样的能力我觉得目前还是有一定的短板的。
泓君：现在机器人公司分很多种，比如说做大脑的，包括这个大脑怎么做，它是Sim-to-Real（仿真到真实）的路径，还是说他用整个物理世界的数据，分很多的流派，还有做本体的，还有做数据的。你觉得Astra出来以后，对哪一类的公司的冲击影响会比较大？
许华哲：先说Sim-to-Real的路线。我一直是认为仿真是没有办法训出最好的机器人的模型和策略的。就很多人说，Astra让仿真重新复活了，因为它做Real-to-Sim，把真实世界迁移到仿真里面这件事变得非常容易。但是它其实只是把它这个3D的外观做得非常好，但是它并没有因为Astra出现，物理引擎变得更准确了。
Astra肯定是利好本体、利好硬件的。Astra出来这两三周，斯坦福已经有一个课题组去把Astra接入一个宇树机器人里面，让它在家里面去做一些家庭任务了。这件事情还是非常快的，就相当于他做了一个论文，两周时间就做出来了。可以看到在Astra有了很强的跨本体泛化能力以后，你的本体如果质量足够好，那可能连接Astra直接可以做一部分事情了，尽管它还只能是demo级。
对于做大脑的公司，我觉得是有两派，一派是相信自己会有一个更快、更好的具身模型，当然可能可以从Astra身上去学到一些，比如数据标注的方式啊、CoT的方式啊等等。我觉得是会带来一些整个行业的增长的，我觉得是一个好事。还有一派，具身模型能力可能不够强，那可能直接放弃掉这件事了，就是说既然Astra好，那我就专注在怎么用它，或者是怎么样把我的东西当做它的一个工具，让Astra来调用我的这些小的东西来完成任务的闭环，那可能更像是一个机器人部署的公司了。
所以我觉得会有一部分公司逐渐走向更偏部署的机器人，有一部分公司会做更小的Astra这样的能力，或者超过Astra能力的模型训练。
泓君：你觉得Astra对整个机器人未来，如何去做研究，会有一些新的启发，或者带来比较大的变化吗？
许华哲：我觉得会带来非常大的变化。第一个是整个模型训练的范式，我觉得大家未来都不会忽略Astra这样的一个范式，所以训模型的人肯定会产生一些新的想法。
另一个就是以前做传统机器人规划的人，那他们现在可能会往agentic，更偏harness这样的方向去做机器人的研究，如何把各种各样的能力组合起来？机器人领域有一个领域叫TAMP（任务与运动规划），它会把规划和控制放在一起，去做长程任务。
那过去这个领域，我觉得是比较古板的，或者说是比较上一个代际的能力，因为它很多时候都用传统的规划手段，去做一步一步的规划。如果我的下一个任务是拧开瓶盖，那我的上一步要确认这个瓶盖首先是关着的，否则的话我不能到下一步。这些事情其实在过去我觉得是相对有一点过时的。但是到了agentic部分，它好像又重新焕发生机了。有了好的模型以后，那又回到规划，可能是变得很重要了。所以我觉得对整个研究肯定也会带来一些变化的。
泓君：您刚刚提到了，比如说一些特别好的机器人大脑的公司，他们可以向Astra学习数据标注的方式，您觉得您从中间获得的启发是什么？
许华哲：我觉得就是要把里面的因果和直觉物理标注清楚，我做这件事情成功是为什么？失败是为什么？我做了这件事情，发生了这个结果，这里面的交互过程需要详细的标注出来，我觉得这个部分是很重要的。
过去的机器人很多标注的时候是只标注一句话描述你做了什么，就相当于只做正向标注，没有背后的原因。比如我没抓住它，是位置偏了，还是东西沾了水太滑了等等。有了这些标注，我们才能真正做一个通用的机器人大脑。
泓君：那这是不是对整个数据标注公司也会有影响？因为我知道现在还蛮多人做机器人数据标注的。
许华哲：我听的传言是OpenAI现在只买不标注的数据，它觉得外面的标注都是错的，他们都是自己内部重新再标注。另一方面，他们可能也是为了保密，不想把它的标注方式透露出去。
所以对数据公司来说会有两种，一种是卖生数据、生肉这种感觉，相当于在卖石油，是一个原材料，主要还是要靠起量来做；还有一部分就是做熟肉、熟数据，那他就需要去了解到最前沿的模型是如何去训练的，如果他有一些这方面的知识，他就可以很自信地说，OK这个是OpenAI的标注方式，这个是Pi的标注方式，你喜欢哪个？我去卖给你，他这个单价可以翻10倍，甚至更高的价格。
泓君：那您觉得它对整个未来机器人研究方式会带来变化吗？因为大家之前是在研究机器人方向的垂直模型嘛，那现在说大模型的能力足够好了，它也可以驱动机器人，虽然它有各种各样的问题，那会不会以后说机器人的训练可能它的底层就还是一个基模，它只用做后训练就可以了。
许华哲：如果我们想要真正意义上部署的话，一个几T的模型肯定是很难去做部署的。如果是远程的话，我觉得是可以解决一部分任务的，但是不能解决所有任务。如果我们都是Astra这种基础模型，我觉得是不太能成为未来所有机器人去使用的模型。一定是会有一个相比于Astra要小得多的一个模型，有相似的能力，一个范式，即使它在云端去做服务、去做host，我觉得也仍然是一个机器人专用的模型。
泓君：您觉得从Astra发布啊到今天，已经大概有三四周的时间。我知道它有机器人画金门大桥的那个demo的时候，很多机器人的创始人是很慌的，后来大家也会自己去试嘛，整个过程您心理上会有一些变化吗？
许华哲：会的。首先它出来的时候，大家刚开始尝试，基本上都是觉得非常的惊艳。最开始尝试当然是让它抓一个东西，然后它很快完成了。后面逐步去尝试更复杂的任务的时候，我们会觉得“机器人的GPT时刻还是GPT”，但是再往深入去看，还是看到很多问题。
第一，它现在双臂的任务是几乎没有办法去做的，比如说在叠衣服这件事情上，Astra是完全叠不了的。所以我们有一个感受就是，在一个语言不太好描述的任务上面，Astra的能力其实并没有很强，所以我们还是看到它在足够灵巧的地方仍然有空间。
第二，我们发现它的总结能力还不是很好，比如说它第一次做一件事，可能会花1小时，第二次做，可能花45分钟，然后我们让它持续不断地去做，它永远都是45分钟，基本上到第三次开始就变成一个常数。不会说像人或者说一些更小的模型，它在充分的训练以后，它的时间会越来越短。所以在这个地方还是可以看到机器人本身的原生模型，还是有很多空间的。
再到现在，我觉得更多的是我们要想系统1、系统2，就是快慢系统，它们之间到底应该怎么结合？
上面的Astra到底是调用，还是说给出一些指令，还是给出什么样的指令作为界面？因为它已经出现了，你不能回避它。但是它自己又有很多问题，你又没有办法直接用它，所以下一个问题是怎么样把机器人公司擅长的东西和Astra这样的模型充分的结合起来，能做出真正通用的机器人。我觉得到现在还是更多从技术角度去想未来怎么做。
泓君：最近还有一个项目也很火，叫Jev，他可能就是很多的是用系统1、系统2、快思考、慢思考的一套理论去做的，你觉得这个对机器人领域的影响大吗？
许华哲：Jev把自己定位成系统1，我觉得跟机器人本身的原生模型是很像的。机器人大家
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Silicate Valley 101
Looking back in September, the AI Circle's heavy news was one by one.
OpenAIpublishGPT-6 Astra, 4 September. This publish report, which does not mention a generic model of robots, drops the head of robots and even renders a group of robots, Benchmark, directly invalid. It's a joke to say: the robotics have waited for so many years for the GPT moment, and it's GPT itself.
Meta Muse supplied each user with a cloud-based computer that hit the top of App Store in the United States in two weeks; the competitor, Inc., increased from 2.5 billion to $10 billion in one month. The individual Agent started a new round of burning money, but they have access to web pages that are being closed down one by one.
In AI for Science, Claude found a new enzyme system like CRISPR in only 21 hours, and the AI circle was excited because AI started doing scientific discoveries, and the biosphere reacted relatively flat, and you wanted to prove it in the lab.
What does researchers feel when a model evolves to the point where it is difficult for ordinary people to feel its evolution? In Silicon Valley 101, we would like to bring together three researchers on the Agent, Robotic and Al Pharmaceutical front lines — Pokee.ai, founder and CEO Chu Chu Chul-qing, founder of the shell-breaking robot and assistant professor at the Qinghua University Institute of Cross-Intelligence, and Wang Zheng, Assistant Professor at the University of Washington College of Computers, to discuss three things: where is the personal Agent commercialization card? How does a robot get hit by a bottom model like Astra? What's the real bottleneck in AI?
Here is a selection of the contents of this dialogue:
泓: Have you ever used this product since Astrapublish? Or what's the most amazing demo we've ever seen online?
Xu Hua Chul: Because I'm a robot, we've measured some of Astra's capabilities in robotic missions. The first thing we see is that it is very resilient. For example, giving it a robotic body to do the right thing, and it'll soon find out that the robot is moving faster, or that the robot's moving space is like, like, it hit the table, and it's gonna say, "This table is so tall, should I lift my hand a little bit higher?" It's actually the ability to transform Cot into a real action that many robot models and LLMs of the past don't have.
The other is its ability to renew and upgrade itself. Because it has such reflections, it can do the same thing once and for all better. And as you can see on the Internet, someone used Astra to operate a robotic arm to draw the Golden Gate Bridge, and it was worse for the first time, but when it was done, it would think of how it could be better, and it could be very authentic after several rounds of it.
We've tried a lot of other things, like flowers, which is a classic robotic mission. We've made it a little more difficult, and it's split in half, a little bit. It'll know at this point, because it's broken, so I'm gonna squeeze this fork, squeeze it a little bit, and stuff it in. It's actually amazing.
CHUNG: It actually understands some of the norms of the physical world.
Xu Hua Chul: I think it's essentially a 3D understanding, because it's also a three-dimensional understanding, but it does help a lot of non-dynamic physical world operations.
CHUNG: I think there's a lot of interesting demo on the Internet that they're making 3D videos with Astra, very fast, and then it's very good to paint.
Xu Hua Chul: This is another astounding robotic capability of Astra, which is to give it a few pictures that can be modelled quickly. The academic community may call it Real-to-Sim. It used to require a lot of video photographs and a lot of manual calibration. But now that you give Astra such a command, it will soon give you the entire laboratory or office environment. So you take this simulation, you can train your robots, and there's an upgrade on the simulation.
Chu Cheol-ching: We're a model company, and Astra, in publish, placed great emphasis on some of their iterative capabilities over post-training, model development, etc. It's not as amazing as it's been, but it does have a more stable improvement over previous versions of the model, both on long-range missions and on an iterative basis. It is able to have a more stable perception of context and the tasks performed earlier than the original model and does not produce too many hallucinations. So it's better in our current experience.
But there may be a few problems: it can be relatively slow, and it may have to do with the current high usage. At the same time, on more long-range tasks, like letting it run for one night, illusions still happen. So it has not been fundamentally resolved, but in general it has improved considerably.
I've got friends with some front-end and design missions. The aesthetics and cognitives at the front end, I think, are quite advanced. It may be during training that there is a special training in design and aesthetic direction.
Z: Bill, this round of Astra's publish, do you think it's a big node compared to the evolution of Coding's capabilities in comparison to the model's impact on industry?
Chu Cheol-ching: Coding alone may not be that big. 4.5 to 4.6, at the time, our sense was an almost paradigm shift, which was supposed to rely on it for short-range Coding, a few file changes, a few simple small projects, and, after the 4.6 model, you basically felt that tool call was almost right, and that it was able to do some long-distance Coding missions, and it could do it autonomously.
And now Astra just says it's a longer mission, the illusions can be lower and the aesthetics better. But it failed to say that I was able to resolve an issue that was completely unsolved before.
Wang: I think it's a big node. We've all switched from previous Claude and others to GPT-6 Astra to do our internal tasks. I'm mostly doing AI for Science, and the core of it is AI for some bimetic applications.
We use Astra in two particularly useful places, one for Computer Use and the other for some long-range tasks.
For its part, Computer Use will use Astra to run software that was developed many years ago in the biomedical field, which is used by default in the field, but they are very difficult to use, interfaces are poorly designed, and there are many terms, so it is very costly for some cross-cutting people to use.
But Astra's Computer Use function can be directly based on a few simple prompts and can work better. And I can learn from it how to use this software, which is, after all, a lot of people have less acceptance of AI tools in the biomedical field. Astra may have adapted these software to some of the online tutorials, some of the user logs that others have previously used, and some of the feedback has been learned in text, so it really understands how to use them.
On the other hand, long-range missions, which are more complex in our biomedicine sector, sometimes lasting between 8 and 24 hours, and Astra doing better. It's slower than some other models, but it's reassuring, suggesting that it might be more deliberate.
We're going to switch to it, and we're going to use some of the faster feedback models in the daytime, and when we're going to let the models run one night, we're going to use Astra to design a longer-range Pipeline or Prompt. This is also a more trustworthy point, some of its manifestations in the field of AI for Science.
CHUNG: I didn't think that the power of Computer Use would help you so much in the AI for Science field.
Wang Zheng: Astra did two biomedical demos this time of publish. One of them is Computer Use demo, a demo in our circle, and it turns out that it was possible to operate a biomedicine tool directly with Astra. It operated a tool called MultiQC, a fairly common web tool. And we started using it, and it was much better than we thought. Actually, the function of Computer Use, I think, is one of the ones that we've had, but you've not really started thinking about how to integrate it into the Pipeline in our research.
Bill, can you explain in one sentence what's Computer Use?
Chu Chu Chulqing: By using models to understand vision, Computer Use decides which part of this agentnext step is going to click on the screen or which text to enter to perform a series of tasks. When it connects to so many dozens of actions, it accomplishs one of the ultimate human goals.
JUNG: The difference between it and API is that you don't have to give an interface, you don't have to give a password, it works like a human being.
Chu Chul-Chu: It sees the screen that humans see, the UI that humans see, and all the visions and interactive interfaces that humans see. These tasks are carried out through the input and output equipment of computers available to humans.
CHUNG: I think that Computer Use had it before. Why did Astra's publish bring Computer Use back on fire?
Chu Cheol-ching: Feels like Computer Use is on fire again. You thought all the tools were going to open the API, all the tools were going to have an interface with permission, and many APs and systems were going to set up.
But lately, whether it's instinc or Meta Muse, it's actually connected to the browser. One of the bigger problems is that there are a lot of third-party APs or Amazons, for example, some shooping APs, and some travel APs, which they are not willing to open up.
And I think it's back to Manus, where Manus started with Browner Use, followed by a lot of others, and you found it not very efficient. Everyone goes back to Coding and API, and now it's like something can't get out of it, and it's going back to Computer Use, so I think it might end up being a mixture.
I have a more interesting idea here, because we're currently a relatively small model of Focus on, for example, that we basically now have publish models around 30B, which can be done on both the end and private calculations. For example, I have an AI box, or I have an AI PC, and it can run this model.
If the future were to say that the Browner Use itself happened on top of your machine, rather than on the cloud, it would change the bottleneck that Browners Use itself could bring to you. Because now the remote Browner Use is using a commercial machine, looking for someone like Bright Data or an IP agent, pretending to be a consumer, to bypass the IP verification. Web sites like Amazon and Expedia usually say that if it's a commercial IP, I won't let you in.
CHUNG: Businesses will block some AI robots and don't want them to access them.
Chu Cheol-sik: Luxury sites will be blocked long ago. If your family's web is your own IP, it won't be blocked. So this could be a fucked-up trend in the future, and this could push a lot of app manufacturers to change their Browner Use mechanism, which means that this Browner Sandbox can happen locally, not in the clouds.
JUNG: Your observation is very interesting, and recently a very hot toc consumer application, Meta Muse, is located as a personal assistant, and it's equivalent to a cloud virtual machine for everyone, and your data is stored in a cloud that belongs to you. According to your theory, it's actually a commercial IP. If it's commercialized in the future, it'll probably get caught by a merchant.
Chu Chu Chul-Ching: Meta Muse, in terms of the virtual setup behind it, was very similar to Manus at the time, and he played a Browner Use because many of its use scenes were linked to travel and shopping, and you had no way of doing this through the API call, basically by the browser. Manus actually took the exact same road and got rid of all the blocks, so Manus went to post-production and now resembles Gensark.
You look at Claude Code, Claude Work, ChatGPTwork, and then you turn to white-collar work efficiency, the central reason being that a lot of things can be done in the first place, but then the giants lose all of them.
泓: The Amazon, for example, has blocked Meta Muse now, and if more websites are added, the fewer pages it may have access to and the less things it can operate.
Chu Chul-ching: But he can't block the IP on your computer, and you can't go to Amazon, so this could be a fucked-up opportunity. But I doubt it's going to come back and forth, and this agent at the C-end, it's probably still not special. Because of Browner Use, the demand for models is actually high. It's hard to pull a phenomenon level in the short term.
CHUNG: Would you look like Meta Muse's long-term development?
Zhu Cheol-ching: I think this wave of consumers will end up in the state of hardware and software integration. I think Meta might have had a chance if she could do something great. And that's why I think that the last thing I want is a combination of land to software and hardware is because the cloud doesn't represent the user itself, and that's why Brownser dropped them off at full black, because once a cloud virtual machine is changed, it can't verify that the person is the user itself.
I think the user's ID or the user's identity, who uses angent, will be a key to this wave toc. If the merchant can confirm that this angent is doing this on behalf of this individual. It's possible that a lot of traders would be willing to do this.
But now, if the cloud-based virtual machine, the cloud-based Brownser, were to be unable to verify this, he would feel that my traffic entrance had been taken away. For example, if I had a user, he was going to visit me, and now he's going to visit me through angent, which might actually be OK for me. But if the original user, who now has 20 angents, whose consumption capacity is the same as it was, and who keep piling up the traffic on my side, it's not worth it to me.
So I think a combination of software and hardware might be an opportunity, if the hardware itself belongs to this user, so he can just confirm that this agent is the user, it's a representational act, and I would be willing to open it to you.
JUNG: Will you use x AI Grok Bot in ToB's workflow?
Chu Chul-ching: I'm more comfortable with Grok Bot myself because it's more proactive. But you can solve these problems with Codex or Claude Code. It's actually workflow, probably less. I'm alone with a lot of people who are really at the B level, like law, medicine, who are really working in traditional businesses, and who are now struggling with how to get out of Microsoft Copilot's muddy. Not even free to use Codex and Claude Code, let alone go to the Grok Bot level. So I think there's a big difference between the whole industry, so people who walk ahead can use anything, and people who walk behind are still struggling.
CHUNG: I've been talking to researchers in the front-line model lately, and I've heard an interesting idea that it's not really going to cost a lot of money to be used by ToC and Consumers. The most expensive place to burn money now is in the core of competition - AI for Science - of Open AI, Anthropic, DeepMind. It's the unsolved, undefined scientific problems, they're going to solve the mathematical questions, or they're going to run some biological experiments like Professor Wang, which is very, very burning token, and it's the biggest internal consumption.
So, for example, the problem of Open AI's Navi-Stokes equation, which calls for a stronger model, with about 10,000 children, sends about 2.7 million messages, consumes about 130 billion tokens. The third party calculated it at a retail price, which was approximately $22 million.
Because now the whole model is evolving, and I think it's not very much for ordinary people. But for cutting-edge scientists, their perception of model capabilities is very high.
Chu Cheol-ching: This place is kind of interested in asking Professor Wang and Professor Wang. I'm doing post training, forcus on small models because we want to do local privatization models. What we can see is that this level of model, after we optimize it, is getting smaller and smaller.
From the point of view of AI for Science, or from the point of view of robotics, how much higher do you think it will be in each generation?
Are we comparing the differences between Opus 4.5 and 4.6, between 4.8 and Fable, or between 5.6 and Astra?
Wang Zheng: My point is that in the field of biomedical science, its AI for Science is indeed a great improvement. But this upgrade is mainly from its coding capacity, a general coding search of literature, and its ability to perform long-range tasks, rather than its ability to improve this biomedical knowledge. Some of the benchmarks are specialized in measuring the Biomedical knowledge level, and Astra has not been upgraded.
It's really hard to use Astra, or Opus, Fable, because they're more and more careful about biomedicine, and it's a serious problem. So we're actually using Kimi K3 or Muse Spark 1.3. I'm not doing a dangerous job, I'm designing a benchmark and so on. So the actual experience with this LLM is getting worse as the capacity grows.
And the other point is that Bill was right, and I agree that on the well-defined AI for Science benchmark, small models are actually getting smaller and smaller. But I personally don't think these benchmarks are as well defined as they are, so the direction you're pursuing, including LLM progress, may not be entirely right.
When these benchmarks are defined, everyone will define a better way for the AI model to do it, for example, with a clear biological data set to predict a clear problem, which is really big for the pharmaceutical discovery that we do. In essence, these current biomedicine benchmarks are a game of data science like Kaggle, so it turns into a machine learning problem in MLE-bench. And astra's ability to coding would make great strides in benchmark.
Xu Hua Chul: I think Astra is a leap up from the point of view of robotics. Neither Opus 4.5, 4.6, nor Fable changed in the past, nor did Astra bring much change to Robotics this time. Because of all the upgrades in the past, almost nothing too good has been done on the Robotics test, and the characterization has not changed.
But the change that Astra now brings is that its entire characterization has changed, and many of the robotics' benchmarks have lapsed. Like RoboDojo, the relative recognition of Robotics, especially the true benchmark, has been the number one. Whatever data it uses, it brings this capability to a leap up.
But on the issue of the little models and the LLM that I've just said, I think we're going to need a small model in the long run, comparable to Astra's ability. The reason is that robotic missions are time-bound, require timely responses and require high-dynamic tasks. For example, I'm going to shake a towel, which Astra can't do because it's too slow. It knows I have to move quickly, but it can only write a code to do, and there's no way it can do it if it goes straight out. In that case, we must need a model for quick decision-making. It's just that we need to add the existing Astra Cot, or its reasoning ability, and 3D understanding.
CHUNG: So do you think the future of robots is going to be the direction of LLM in planning, and small models are going to take control.
Hsu Hua Chul: The hierarchy has always been in agreement, but where is the layered knife? There has always been debate on how to stratification. So, for example, when it was generally believed that VLM was a mission planning level, the mission planning level probably came, and I told you I'd eat tomatoes and eggs, and the VLM plan might be to open the refrigerator door, take out some tomatoes, take out the eggs, etc., and then what happens if it's a little model or a body model, and it's going to do it? What do you do with eggs?
But the biggest change in Astra this time is that it can now do low-level movement control, which is the equivalent of the order it gave you, which is to stretch two centimetres to the right, or put its hand on the handle of the fridge door, and if it is not caught, he says that you lean forward one more time... so it's the equivalent of pressing down this layer.
Turns out everyone's borders are clear, but now it says I'm strong enough to press down a little bit, except slow. It's more like a reflex-like wire puppet. There must have been some voices saying, "Is that a future top model that can crush indefinitely?" Is it a complete-to-end VLM when it comes to the end? Astra directly controls everything.
My personal point is, I don't think it makes any sense. The first is clearly speed, and all the controls you see on the Internet are 16 or 24 times. Second, I think that there's still very little understanding of physics in Astra today, especially for humans called intuition, or the part of intuition physics.
I could touch an object with my own hands if there was water that could spill, if glass could break, something that Astra could understand if it went through many language reasonings, but if we wanted it to do a task that immediately had that understanding, it might not have been effective enough. This place may still need to be like a world model, or like a fast model, to have some intuitive reaction.
CHUNG: Did you just mention its slowness, its slowness because of what? Is its model too big?
Xu Hua Chul: The model is definitely one of the reasons why it's a long CoT, although it's a encrypted CoT, and the CoT that comes out is a fake, I understand, but its internal CoT is a very long one, so it makes every step of the decision-making, and it's very redundant, especially when it comes to simple tasks, and we observe that complex tasks and simple tasks, which are basically similar in length, mean that complex missions are smart, right? But when it comes to simple tasks, it becomes an extra cost, making it look very stupid.
Prof. Xu, I know that your company has a demo with chopsticks and balls. What's the difference between your own models and Astra?
Hsu Hua Chul: With our own model, first, it's much faster in time; secondly, it's a continuous attempt at catching, and there's not much logical thinking in the process. I caught a failure, and I could try from another angle.
But if it were Astra, after every step it had taken, it had to analyse what happened, and for that reason it had to give itself instructions and what adjustments I would have to make. This process seems more like this human process of thinking, but it leads to a slowness; secondly, it sometimes leads to stalemates, it sometimes fails to understand why this is happening, or it does not have the means to move forward once it is analysed wrong.
Another point we see is that if we go to grab an object on the table and spin this thing, Astra actually does it relatively weakly, but it can do it, and it tries repeatedly to turn it around, so I think there's still a short board.
CHUNG: Now robotics have a lot of different kinds of things, such as brain making, including how the brain does it, is it the Sim-to-Real path, or does he use data from the entire physical world, a lot of genre, a lot of genre, a lot of genre, and a lot of data. What kind of company do you think Astra's impact will be greater when she comes out?
Xu Hua Chul: First, the Sim-to-Real route. I always thought that there was really no way to get the best robot models and strategies. For a lot of people, Astra brought the simulation back to life because it became Real-to-Sim and it became very easy to move the real world into the simulation. But it's just doing a very good look of its 3D, but it's not getting more accurate with Astra.
Astra's got to be real and hardware. Two or three weeks after Astra came out, Stanford already has a team of people who are going to get Astra into a wild tree robot and let it do some family work at home. The event was still very quick, and it was equivalent to a paper he had made in two weeks. And you can see that with Astra's strong cross-substantiation, your body, if it's of good quality, could be connected to Astra to do some of the things directly, though it can only be demo level.
I think there are two kinds of companies that believe that they have a faster and better body model, and of course they can learn from Astra, the way data are labeled, the way of Cot. I think it's going to lead to some whole industry growth, and I think it's a good thing. There's another type of body model that may not be strong enough, and it might just get out of hand, which means that if Astra is good, I'm focusing on how to use it, or how to use my stuff as a tool to get Astra to use my little stuff to do the job in a closed ring, which may be more like a robot-deployed company.
So I think there's gonna be some companies moving towards more deployed robots, some companies going to do smaller Astra-like capabilities, or model training beyond Astra's capabilities.
CHUNG: Do you think that Astra's research into the future of the robot as a whole will provide some new insight or bring about greater change?
Hsu Hua Chul: I think it's going to change very much. The first is the paradigm of model training as a whole, and I don't think you're going to lose sight of a paradigm like Astra in the future, so there's definitely something new coming from the trainers.
The other is the former conventional robot planners, who may now be studying robotics in the direction of agentics, more than harness, how to combine all kinds of capabilities? There is an area of robotics called TAMP, which brings together planning and control for long-range tasks.
In the past, this area was, I think, more rigid, or more intergenerational, because it used traditional planning tools to plan step by step. If my next task is to break the cap, then my last step is to make sure that it's first closed, otherwise I can't go to next step. These things were, in fact, kind of outmoded in the past. But on the part of angentic, it seems to be rebooting. With a good model, it's probably going back to planning. So I'm sure the whole study will make a difference.
JUNG: You just mentioned, for example, some very good robotic brain companies that can learn from Astra how to label data. What do you think you're getting from the middle?
Hsu Hua Chul: I think it's about making clear what causes and consequences and what instinctive physics are inside. Why am I doing this? Why fail? I did this, and this happened, and I think this part of the interactive process is important.
Many of the robots of the past were labeled with a single sentence describing what you had done, which was a positive note, without the reasons behind it. Like if I didn't catch it, if it was in the wrong place, or if it was too watery. With these markers, we can actually make a universal robotic brain.
CHUNG: Does this not affect the entire data label? Because I know there's a lot of people doing robotic data.
Xu Hua Chul: The rumor I hear is that Open AI now only buys unlabeled data, and it feels that the outside labels are wrong, and they're re-marked from within. On the other hand, they may also be seeking to maintain confidentiality and do not want to disclose its labelling.
So for the data company, there would be two kinds of feeling, like sales of raw data and raw meat, which would be like selling oil, a raw material, and whether to do it primarily by starting up; and part of it would be to make ripe meat and ripe data, so he would need to know how the front-line model was trained, and if he had some knowledge about it, he would be confident that OK would say that this is Open AI's label, this is Pi's label, which you like. I'll sell it to you. He can double the unit price ten times, even higher.
CHUNG: Do you think it will change the way the robot is studied in the future? Because you were looking at vertical models of robot direction, and now it's good enough to say that LLM can drive robots, even though it has a variety of problems, would it be possible to say that robots' training may be at the bottom of a model, and that it can only be done after training.
Xu Hua Chul: If we want to deploy in the real sense, it must be difficult to deploy a few Ts. If it is remote, I think some tasks can be solved, but not all. If we're all models like Astra, I don't think it's going to be a good model for all robots in the future. There must be a much smaller model than Astra, with similar capabilities, a paradigm, even if it's in the clouds doing service, doing host, and I think it's still a robotic model.
CHUNG: You think it's been about three or four weeks since Astrapublish. I know that when it had a robot that painted the Demo on the Golden Gate Bridge, the founders of many robots were scared, and then people would try it themselves.
Xu Hua Chul: Yes. First of all, when it came out, you all started trying, and it was basically amazing. The first attempt, of course, was to get it to grab something, and then it was done. And then, as we step back to the more complex tasks, we think, "The GPT of robots is a GPT at the moment," but look at it in depth and see a lot of problems.
First, there is little that it can do with its arms now, for example in the case of folding clothes, and Astra can't do it at all. So we have the feeling that Astra's ability is not really strong on a task that is not well described in a language, so we still see that there is room for it in a smart enough place.
Secondly, we find that its summing-up capacity is not very good, for example, if it takes one hour for the first time, if it does the second time, if it takes 45 minutes, and then we let it go on and on, and it's always 45 minutes, basically starting with a constant. Not to say people, or smaller models, it'll be shorter and shorter after it's fully trained. So there's still a lot of space in this place to see the prototype of the robot itself.
And now, I think it's more about how we want systems one, two, fast and slow, how they fit together.
Did Astra call, or did he give some instructions, or what kind of instructions as an interface? Because it's already there, you can't avoid it. But it's got a lot of problems, and you can't use it directly, so the next question is how to fully combine what robots are good at with models like Astra and make really universal robots. I think there's more to be done now from a technical point of view.
CHUNG: There's also a recent project called Jev, which is probably a lot of theory using system 1, system 2, quick thinking, slow thinking. Do you think this has a great impact on robotics?
Xuhua Chul: Jev has positioned himself into System 1, and I think it's very similar to the original model of the robot itself. The robots.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.